Report vulnerabilities through GitHub Security Advisories. Do not open a public issue for a security report.

This repository ships a pattern specification, a directed workflow, and a read-only linter. It runs no service and stores no data. Its attack surface is therefore mostly what it tells other people to build.

Security issues here include:

• Guidance that produces an unsafe viewer. The pattern's first law is that the observer must never be able to affect the observed. Any phase instruction, template, or example that would lead an implementer to generate a viewer capable of signalling, writing to, locking, or killing the job it watches is a security defect in this repository, not merely a bug.
• watchctl performing any write, execution, or network access. The tool reads files and prints findings. It must never execute a script it is linting, follow a symlink out of the tree it was pointed at, or open a socket. A linter that violated its own first law would be worthless.
• A law check that produces a false pass on a genuinely dangerous construct. A missed mutation in a viewer is more serious than a false positive, because the pragma system means a clean report is read as an assurance.
• Pragma bypass. watchctl:allow requires a reason by design. A way to suppress a law without recording why is a defect.
• Guidance that leaks secrets onto a screen. A viewer renders whatever the job wrote. Phase instructions that would surface credentials, tokens, or personal data into a terminal — or into a pasted NOCOLOR=1 snapshot — are in scope.

Not in scope:

• Vulnerabilities in a watcher you generated. Those belong to your repository.
• The example source repositories (leather, adherence-suite). Report those to the repositories themselves.

The linter is a deliberate subset. L6, L11, L16 and L17 are judgment, not syntax, and watchctl does not check them. A clean watchctl lint means the mechanically-checkable laws hold — it is not a certification that a viewer is safe. This is stated in the tool's own documentation and must stay stated.

Pragmas are honoured on trust. watchctl:allow L1 <reason> suppresses a finding on the strength of a human-written justification. The reason is mandatory so it can be reviewed, but nothing verifies it.

The latest tagged release. Fixes are not backported.